Clustered errors

[lkonya]

Has anybody estimated clustered standard errors with Eviews? I know that it can be done with Stata, but my preferred option would be Eviews.

Thanks

[EViews Gareth]

EViews doesn't have this built in at the moment.

[EViews Glenn]

Not quite. Set it up as a panel with your cluster as the cross-section identifier. Then there are various options for grouped robust errors on the cross-section dimension.

[msr]

Hi!

In fact, there seem to be several methods.

However, sofar I cannot see, which of the methods allows for
* non-identivcally distributed disturbance terms and
* correlations of disturbance terms within a cluster.

These two features seem o be of particular relevance in
empircal corporate finance (and seem to be standard for
researchers using stata).

Any help is highly appreciated!

Best regards...msr

[EViews Glenn]

If you define your panel with the cluster variable identifying the cross-section dimension, the "White cross-section" covariance option will estimate using what Arellano (Panel Data Econometrics, 2003, p. 18) terms the fixed T and large N robust standard error that is robust to heteroskedasticity and serial correlation (within cluster) of arbitrary form.

[lkonya]

Thanks very much. I'll try to apply the recommended procedure.

[msr]

Dear Glenn

Sorry for bothering you one more, but I still have problems with these options.

In the help file it says that

Using white period standard errors & covariance gives SEs that are

…robust to arbitrary serial correlation and time-varying variances in the disturbances. [...] The White period robust coefficient variance estimator is designed to accommodate arbitrary serial correlation and time-varying variances in the disturbances. 

Using white cross-section standard errors & covariance gives SEs that are

…cross-equation (contemporaneous) correlation as well as different error variances in each cross-section.

Given your answer, it seems that the latter is what I am looking for. Now, using the latter gives me (sometimes) suprisingly small SEs (in a panel data set with large N and small T).

Again, what I am looking is a method that allows (in a panel data set with large N [firms] and small T [periods]) for
* cross-section heteroskedasticity, i.e. for a different residual variance for each cross section
* arbitrary correlation within a cluster [i.e. within a firm]

Any help is highly appreciated!

Best regards...msr

[EViews Glenn]

Sorry about the confusion, it should be White period, which allows for arbitrary period correlation structures.

See Arellano (2003). Panel Data Econometrics, Section 2.3.1. p. 18. Requires asymptotics in N and fixed T.

[julia]

Dear Glenn

I have read the earlier posts, but I am still confused and I don't know whether to use White Priod or White Cross-Section SE & covariance.

I do not do panel analysis, but - as you suggested in an earlier post - I have set up the data as a panel with the cluster variable [issuer] as the cross-section dimension.

I would like to have heteroskedasticity-robust standard errors clustered at the issuer-level, in order to allow for correlation in standard errors that is specific to an issuer.

Thanks for your help and patience!

Julia

[EViews Glenn]

If your workfile is structured as a panel with "issuer" as the cross-section identifier and you want to allow correlation for observations within issuer, but not across issuers, then the White - period will allow for between-period correlations (i.e., clustered by issuer).

The unfortunate thing is that there are two different conventions for naming these things which is why it's all a bit confusing. Let me see if I can provide some background to help clarify things.

The original econometric literature on SUR type models was based on macro models with different equations for each cross-section (in the typical case, country) where it was believed that observations were contemporaneously correlated. Roughly speaking this error variance structure was given the name "cross-sectionally correlated" to indicate that cross-sections were correlated. Hence the name of cross-section SUR for models which allowed for contemporaneous correlation but no between-period correlation. Note that in this convention, the "cross-section" nomenclature references the dimension that has correlation.

A more recent statistics literature would refer to this data structure as clustered "by period", since for each period the observations are correlated, but there is no correlation between periods. In this world, the "period" references the unit for which there is within-correlation, not the units across which there is correlation.

When we introduced the two different forms of SUR and White covariances we tried to match the terminology of the original econometrics literature (for better or worse). Thus, "White - Period", may be thought of as a White estimator where we assume that there may be between-period correlation, where there is no correlation across cross-sections. These data would be termed clustered by cross-section in the second literature.

Now personally, I find the clustering nomenclature to be more natural, but that may not be the case for macro time series people. We may, in the future, switch the labeling over, but would do so in a way that preserved backward compatibility.

I hope that this clarifies things.

[julia]

Thank you very much! This helped me a lot.
Regards,
Julia

[mangalenka]

If you define your panel with the cluster variable identifying the cross-section dimension, the "White cross-section" covariance option will estimate using what Arellano (Panel Data Econometrics, 2003, p. 18) terms the fixed T and large N robust standard error that is robust to heteroskedasticity and serial correlation (within cluster) of arbitrary form.

what if the cluster variable is Country but the way my file is set up Firm is the main cross-section identifier and Country is just one of the many characteristics of the Firm eg the Firm is based in the USA. How can I manipulate the file to make Country the cluster variable so I can use the "White cross-section" covariance option?

Thank you!

[EViews Glenn]

As I noted earlier, since we don't cluster arbitrarily, you'll have to restructure as an unbalanced panel on country for the purpose of computing country clustered errors.

[MalaMi]

Hi, I want to estimate a cross-sectional regression of individual subject’s income on several individual parameters. However these data were obtained through several experimental sessions. My professor suggested that there could be inter-session correlation between the outcomes, which I should account for. Can I organize my data, although it’s not a panel, as a panel and while estimating it as a panel regression correct for clustered errors? There are 8 subjects in each of the 5 sessions. Each subject participated only in one experimental session. I created a balanced panel where I set “session” as a cross-identifier and “subject” as a time variable. I am not familiar with panel-data regression and I’m very much concerned that setting subject as time variable is wrong? Thank you in advance!

[EViews Glenn]

Yes to the basic idea, though with only 8 subjects, the asymptotics for the entire approach are a bit iffy.